tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-15 04:33:06 -04:00
Code Issues Packages Projects Releases Wiki Activity

kv-cache : fix seq_rm with seq_id == -1 (#15226)

Browse Source 
* kv-cache : fix seq_rm with seq_id == -1

ggml-ci

* cont : iterate over streams

ggml-ci
This commit is contained in:
Georgi Gerganov
2025-08-11 13:58:24 +03:00
committed by GitHub
parent cd3069dfcb
commit 228f724d9c
1 changed files with 30 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

48
src/llama-kv-cache-unified.cpp
View File

@@ -223,12 +223,7 @@ void llama_kv_cache_unified::clear(bool data) {
} }
bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) { bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()); GGML_ASSERT(seq_id == -1 || (seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()));
auto & cells = v_cells[seq_to_stream[seq_id]];
auto & head = v_heads[seq_to_stream[seq_id]];
uint32_t new_head = cells.size();
if (p0 < 0) { if (p0 < 0) {
p0 = 0; p0 = 0;
@@ -239,6 +234,11 @@ bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
} }
if (seq_id >= 0) { if (seq_id >= 0) {
auto & cells = v_cells[seq_to_stream[seq_id]];
auto & head = v_heads[seq_to_stream[seq_id]];
uint32_t new_head = cells.size();
for (uint32_t i = 0; i < cells.size(); ++i) { for (uint32_t i = 0; i < cells.size(); ++i) {
if (!cells.pos_in(i, p0, p1)) { if (!cells.pos_in(i, p0, p1)) {
continue; continue;
@@ -250,26 +250,38 @@ bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
} }
} }
} }
// If we freed up a slot, set head to it so searching can start there.
if (new_head != cells.size() && new_head < head) {
head = new_head;
}
} else { } else {
// match any sequence // match any sequence
for (uint32_t i = 0; i < cells.size(); ++i) { for (uint32_t s = 0; s < n_stream; ++s) {
if (!cells.pos_in(i, p0, p1)) { auto & cells = v_cells[s];
continue; auto & head = v_heads[s];
uint32_t new_head = cells.size();
for (uint32_t i = 0; i < cells.size(); ++i) {
if (!cells.pos_in(i, p0, p1)) {
continue;
}
cells.rm(i);
if (new_head == cells.size()) {
new_head = i;
}
} }
cells.rm(i); // If we freed up a slot, set head to it so searching can start there.
if (new_head != cells.size() && new_head < head) {
if (new_head == cells.size()) { head = new_head;
new_head = i;
} }
} }
} }
// If we freed up a slot, set head to it so searching can start there.
if (new_head != cells.size() && new_head < head) {
head = new_head;
}
return true; return true;
} }